Introduction: Model-based machine learning offers alternatives to the classical feature extraction techniques and also to the general-purpose black-box deep learning methods. Model-based neural networks support the direct incorporation of the domain knowledge into the network architecture, providing optimized representation learning, better computational properties, and explainability, which emphasizes their importance in biomedical applications. In this work, we propose a novel model-based neural network architecture by unfolding Variable Projection (VP) iterations in order to automatically classify ECG arrhythmia.
Methodology: VP is a general framework for nonlinear optimization problems that has already been successfully utilized in biomedical signal processing, with applications like adaptive feature extraction, segmentation, and classification of EEG and ECG signals. Here, we utilize the deep unfolding technique for VP, and we propose a network architecture where the layers are designed to resemble a single iteration of VP combined with multi-layer perceptrons. We investigated VP with adaptive Hermite function system, where the trainable parameters (translation and dilation) are directly related to the morphology of the ECG heartbeats, thus providing interpretable behaviour.
Results: The proposed neural network architecture provides a model-based end-to-end framework for ECG arrhythmia classification. We provide the mathematical and computational background of the unfolding process, and the evaluation of the proposed method on the benchmark MIT-BIH Arrhythmia Database, where 3 standard AAMI arrhythmia classes were considered (normal, supraventricular, and ventricular heartbeats), employing the subject-oriented evaluation scheme. We achieved overall accuracy over 94%, which is a competitive performance compared to the state-of-the-art.